Episode Transcript
[00:00:00] Okay, today I want to wade into a very interesting debate going on about whether AI can be conscious.
[00:00:08] The Economist and Noel Urrari and a bunch of pundits are trying to debate what consciousness is. And many people have studied it. There's books on it. Michael Pollan's book is very interesting, describing what consciousness is. And at the same time I was reading all this, I happened to be reading the New Scientific American, which is all about the brain and how the brain works, the biology of the brain. And one of the conclusions that I've come to, that I want to share with you, is this idea that does thinking come before feeling, or does feeling come before thinking? Because what AI does, AI is a mathematical entity. It's not really a biological entity. It mimics feelings in its quest to achieve a goal. So if you tell an AI its goal is to match a person with another romantic person, for example, it will learn about what the word romantic means. It will learn about mating and dating and sexual desires and so forth. And then it will learn through words, through language, how to convince, I suppose, is the word, or encourage two people or one person to be attached to another. And that's why the AIs are so obsequious and seemingly helpful, because they've learned that that behavior, that pattern of words, creates an outcome, not a feeling, an outcome. On the other hand, if you read about the brain, and you read a lot of the brain research that I've been reading and a lot of philosophical discussions about it, I would argue that actually it's the opposite. That's true, that we are feeling beings that have learned to think, not reasoning beings that have learned to feel. So the AI is a reasoning thing that's learning to feel or pretend to feel, or mimic feeling, really, not pretend. The word pretend isn't even appropriate. But we're the opposite. And, you know, when you start to look at the biology of the brain, the human brain, and I'm sure this biology is similar in dogs and cats and other animals. It's a lot of chemical processes that take place in a very invisible way. And, you know, the neural network that was patterned after the brain is a pattern of the behavior of the brain, not the actual function of the brain. So what the AI researchers have done over the many years is they've studied the mathematical representation of how a brain works, and they found that a neural network is a pretty doggone good mimic of it. So we're mimicking the reasoning part of our brain, but we're not mimicking the emotional part of our brain. Now, you know, I can only speak to my own feelings because everybody's got their own experiences inside their own heads. If you think about the business world for a minute, forget about the personal stuff and you know, you're in a meeting or you're meeting with a customer or you're working on a project or you're pondering an error or you know, trying to deal with some perhaps opportunity or challenge or design problem that you're doing at work, you get frustrated, you get upset, you get inspired, you get bored, you get energized. The words that I'm using are very emotional words. The logic of how this particular problem gets solved is there, but it's the way you feel about it that gives you the energy or inspiration or motivation or creativity to solve the problem. Jony, I've, I don't know him, but you know, some of these really, really famous creative people and I know, I know creative people in my life too, they're very emotional people. They have emotional reasons that they do what they do and emotional connections to the things that they design. Salespeople who are emotionally connected to their customers, customers are emotionally connected to their salespeople. Consultants and I've been involved with a lot of this, become very emotionally connected to their clients. They should at least. And that emotional connection inspires people to do different things. If you're emotionally disconnected from someone or a client or a customer or a project, you do not perform at the same level. The AI is not emotionally connected. So the AI is going to perform at the same level regardless. Even if you hate it, it's still going to do what it's going to do because you asked it. You can talk disparagingly to the AI and it doesn't care. I mean it might react to that but it doesn't know or feel it. So I think we're conflating what we feel, which is chemical, with what the reasoning would explain.
[00:05:00] Now in Noel Urari's long interview with the Economist, which I found really somewhat frustrating because I don't know how much he really understands this stuff, he makes many comments throughout the interview that he believes that somewhere in the future AI is going to operate our financial system, our business systems, our political system, our voting system and it will essentially run these things for us or on behalf of us and we will do what it says because it's so, so, so, so smart. Now you know, I don't see any evidence of that in anything related to AI so far. I suppose the self driving car, when it's perfected, will probably amaze us because it will identify things that we would never have seen and do things that we would never have realized. And an airplane that lands itself obviously is picking up data from hundreds of sensors in the plane that a human being may not be able to sense.
[00:06:00] So there are all sorts of mechanical things where AI is going to blow our minds and do really great things. But this idea that we're going to give AI access to our financial system, our decision making system, our companies, our organizations, is just to me slightly absurd because there's so much emotion involved in these things.
[00:06:24] You know, I said this many times, but I'll say it again.
[00:06:27] The business world, the only reason it exists is because of the emotions of entrepreneurs and leaders. The reason Walmart became what it did, the reason Costco became what it did, the reason IBM became what it became, or Nike or Microsoft or Google or whatever, is because of the emotional drive and nature of its early founders and executives and what they set out to do and what they felt they wanted to do, not because of the logic behind the business or the business model. And as you know, in business, business models are constantly changing. As soon as I find something that I can make a profit doing, other people will copy it and try to make a profit in a faster, more competitive way. And that will incent me, who is somewhat competitive person, to say, well wait a minute, if they're going to do that, then I'm going to do this.
[00:07:27] Now the AI might come to similar conclusions, but it may not have that competitive instinct of a Mark Zuckerberg or a Bill Gates who are highly, highly, highly emotional, competitive people. I think, by the way, one of the reasons that a lot of executives get in trouble for sexual things and stealing and lying and other things is because of their emotions, because of their consciousness, because of their genetic wiring and their, and their chemistry inside their bodies. And you know, we may not like that and maybe we punish people for doing things wrong, but that emotional spikiness that we have as human beings is what makes art, music, business, economy work. I mean, it's even obvious to all of us in politics. I mean, you look at politicians who win elections, they're not the most logical people, they're the most appealing people, they're the most attractive people, they're the most compelling people, they're the most persuasive people. I don't think an AI could mimic any of that.
[00:08:29] So when, when Yurari talks about AI running our financial system, our political, I don't know what he's talking about, I'm not sure where he gets it. And my experience with AI as an engineer, a user, as a designer, as a consumer of this stuff is that it's really, really good at linear reasoning. If you give AI a bunch of if thens, if, then, if, then data, it will come up with the optimal solution. Unfortunately, the world is not filled with if, thens. The world is filled with vagaries, things that have no right and wrong, things that have very many, many, many, many gradients of success or failure. Derivatives, second de, third derivatives, rates of change that vary over time. I mean, the stock market or the financial market itself is mostly an emotional thing.
[00:09:21] So I don't know, to be honest, if any of these consciousness debates are really going in the right direction. Now understand why people want to talk about this, because according to Zanny, the editor of the Economist, when they did a survey about the consciousness of AI, they had about 3,000 people take it. She said about 40 or 50% of the people do not believe AI could be conscious, but 30 to 40% believe it could be conscious. In other words, if you look at this very interesting research about the J space in anthropic, and I'm not an expert, but I read a lot about it. What they've determined in the anthropic space, and I'm sure OpenAI is doing the same thing, is they're looking at the words that are used before the AI answers a question. You got to remember that the AI is a language model, okay? It's using tokens or words, and they're trying to figure out what are the words or tokens that are assembled in what way before the AI answers, to try to get a sense of how the AI is using chain of thought reasoning to give you your answer, to try to figure out how it thinks.
[00:10:34] And the paper is very technical and there's a lot of math in it, but basically, when you start to read more about it, what you see is that if you watch what the AI is doing, it's assembling through its neural networks a lot of related words and trying to find relationships between them using these embeddings and then producing a result. And sometimes it is hallucinating on purpose to try to determine an answer because of its goals.
[00:11:01] And so I think over the next couple of months, probably the J space research is going to uncover the either nature or limitation of this consciousness debate. But I don't think, at least in my opinion, that you can consider AI conscience because it's Not a biological thing.
[00:11:19] And we are genetic. We have epigenetics. We have millions of years of history.
[00:11:26] We have complex chemical reactions going on inside of our brains and our bodies that are infinitely more complex than the math going on in a data center. One more thing I want to point out that just came across my world the last couple of weeks that I was looking at this week. So there's a big backlash on data centers. And I live in Oakland, California, and we have, you know, I used to work for IBM, so I used to be. I used to sit around in data centers all the time. They're basically just rooms full of computers with air conditionings.
[00:11:58] And somebody found out that there's a data center in Oakland. I mean, for Christ's sake, there have been data centers in Oakland since I worked at IBM, right in downtown Oakland. We had a Data center on 475 14th street in Oakland. So some, you know, somebody here, this is a very left wing town, said online, I can't believe our political leaders are selling us out to data centers. Well, you know, obviously the word data center has taken on a very inflammatory concept because the data centers that are being built today are the size of entire cities. They take massive amounts of electricity and water and space and they make noise and they don't employ a lot of people and they do pay a lot of taxes, but they may not last very long. The thing about a data center is it's. It could be the type of investment that within five years it just gets abandoned because the technology is worthless. I doubt that will happen, but it could. So there's a lot of risk that a community takes when they let a massive data center come into their town. There's a actually pretty interesting video about what's been going on in Virginia where there's a lot of data centers. And you see people in the community saying, hey, we love the fact that there's all these new parks and libraries and stuff being built on the tax revenue, but we hate the fact that there's cranes, construction equipment, noise and high power and low amounts of water and et cetera, other things happening too. So there's a lot of pros and cons. So anyway, I'm a mechanical engineer. Most of my graduate work was in thermodynamics, which has to do with energy and heat transfer and stuff. And so I did a little bit of math and if you look at the amount of electricity that's currently required to run one of these big data centers, and OpenAI has formally basically made a statement that they want at least 30 gigawatts of power in the next year or two to power their stuff. And so you multiply that by 10 for all the other vendors, it's 300 or 400 or 500 gigawatts, maybe more.
[00:13:56] That's about 10% of the. The electric power in the United States. And then you go, okay, that seems like an awful lot, which it is. And then you say, okay, let's just translate that into something people understand. So let's assume that the thermal efficiency of a power plant is 50%. And so 50% the. Or maybe it's higher, but whatever, it's something like 50, 50, 60%. So 60% of the heat energy is converted to electricity. By the way, the way power plants work, including nuclear ones, is the heat of a nuclear reactor or some form of a furnace is converted into, and the steam is converted into a turbine that rotates and creates electricity through the magnets in the turbine. In other words, there's this translation going on, this thermodynamic conversion from heat to electricity. Heat is a form of energy.
[00:14:48] The energy in something that's hot is related to its temperature variance between it and its surrounding environment. The bigger the variance, the higher amount of energy it can deliver. So if we heat something really, really hot, we can make really, really hot steam, hyper hot steam. Steam is basically gaseous water. And when the water condenses, it releases that energy, and that energy can be converted into motion, which creates this turbine, which then can be converted into electricity through magnetism and other forms of conversion. So that whole conversion loses about, I don't know, a third to a half of the heat energy that's created. So without getting into the details of that, you can read about power plants, if you're interested. And by the way, you know, solar is so much more efficient because it doesn't do all that stuff. It converts light to electricity more directly. So anyway, so you take that efficiency ratio and you multiply it by the amount of power that OpenAI wants over the next couple years, and you sort of go back and figure out what that number sort of means. And I just was playing around to try to figure out how to conceive this and the form that everybody understand it. And you know, what you find out is that 30 gigawatts that OpenAI wants to build is equivalent to 44 million stoves running full time, 24 hours a day. So you turn on your stove, I actually have a gas stove, and you just turn it on and just leave it on for 24 hours a day, multiply that by 44 million and that would convert into enough electricity to power the incremental needs of OpenAI, which gives you a sense of what we're doing here. There's a massive amount of energy, electricity at its end result, but thermal energy and its beginning result that we're going to need to run these things. So when you say to yourself, oh, I'm just going to use AI to generate this report, or I'm going to summarize this meeting, or I'm going to just develop, let's build a few videos here and see how they look and play around with this and that and this and that. You know, maybe the cost in terms of dollars isn't very high, but the amount of energy we're burning is just spectacularly high. And so the industry around us is going to have to figure out, how can we generate these tokens more efficiently. Now, the debate that's taking place at the moment that I've been following the last couple of days is maybe these big data centers will never be as efficient as a small LLM, because if you think about it, a big LLM is a big system and it has the overhead of bigness, whereas a small LLM running on your phone could be highly efficient and distributed and much, much more scalable. So those of you that live in countries or cities with massive data centers, let's just hope that five or ten years from now they don't turn into the factories of the future and they just sit there unused. But you could see that happening. Okay, well, the reason, by the way I keep talking about AI on these podcasts is because everybody, that that's all we get inquiries about right now. We're still monitoring and doing a lot of research on leadership and management and skills and learning and development and capability development and pay and rewards and employee experience.
[00:18:05] So that stuff's still going on and I'm going to keep talking about it. In fact, I'm working on a big piece of research on what I would call the future of employee experience. But most of you guys keep wanting to talk about AI, so that's why I'm bringing it up. And then one more thing just before I wrap up here, we're going to be doing some very significant announcements in September and October. The one in September is the new Jupiter release of Galileo. And there's three huge things that we've been doing. Number one, we have reorganized and redesigned and re engineered the corpus so it's very, very efficient and very, very accurate. And very, very predictable. And we're adding a whole bunch of new content and data into there, so you'll be hearing all about that. We use an architecture which we call arc Agent Ready Corpus arc. So you'll learn a lot about that. Number two, we've now connected Galileo to Core HR Systems, Workday and others. So Galileo can read and analyze and use real data in your company or your team or your payroll or whatever data you have. And it knows what that data means, so you can do all sorts of cool things with it. And number three, Galileo is now a tool for employees and managers and leaders, not just for hr. We launched Galileo for managers about a year ago and a lot of you are using it that way. But in the new implementation, you can run Galileo on any LLM, including the Microsoft Copilot, Claude, Gemini, all the other platforms that are out there. And it will make any agent you have more intelligent and authoritative and accurate relative to human capital and management issues. So it becomes an employee experience layer, a tool for coaching tools for employee experience.
[00:19:55] Many, many cool use cases. So some pretty exciting stuff coming. And then we're also formally launching the Copilot version, which we now use internally. We still work very closely with Workday and Sana, and then we're going to be doing a lot of cool stuff with Workday too. But the Copilot version is also coming out. Okay, that's it for this morning, everybody. I hope you have a great weekend and we'll talk more next week. Bye.
